FastDesign

Fast RNA Design via Motif-level Divide-Conquer-Combine and Structure-level Rival Search

Installation

1. Clone the repository:

   git clone https://github.com/shanry/FastDesign.git
   cd FastDesign

2. Install dependencies: This project is developed using Python 3.11, and should also work for other Python versions.

   pip install -r requirements.txt

3. Optional: Enable Rival Search To enable structure-level rival search for boosting MFE-based metrics, download and build the RNA-Undesign repository.

   Then, set the PATH_UNDESIGN environment variable:

   export PATH_UNDESIGN=/path/to/RNA-Undesign

Testing

pytest main.py -s

Quick Start

You can run FastDesign directly from the command line.

Fast Version:

echo ".((((((((((..((((......((((((((.......)))))))).(((..(((((.....)))).)..)))))))....))))))))))" | python main.py --online

Full Version:

echo ".((((((((((..((((......((((((((.......)))))))).(((..(((((.....)))).)..)))))))....))))))))))" | python main.py --online --poststep 2500

Main Parameters

The primary command-line parameters are:

• --path:
  Path to the input file containing target RNA secondary structures, one structure per line.
  Examples: data/rnasolo764.txt, data/eterna100.txt.

• --step:
  Number of optimization steps for leaf-node (motif-level) design. Default: 5000.

• --poststep:
  Number of optimization steps for root-node (full-structure) refinement. A flexible parameter that balances quality and efficiency. Default values:
  • 0 for the fast version
  • 2500 for the full version

• --repeat:
  Number of repeated experiments to run. Default: 1.

• --worker_count:
  Number of CPU cores to use for parallel execution.

• --batch_size:
  Number of structures processed in parallel per batch. Must be ≥ worker_count.

• --k_prune:
  Beam size for cubic pruning. For each node, the top k_prune candidate designs are retained during search.

• --motif_path:
  Path to easy-to-design motifs.

Experiment Replication

The replicatable results (RNAsolo_full.zip, RNAsolo_fast.zip, Eterna100_full.zip, Eterna100_fast.zip, 16SMFE_fast.zip) are available on Google Drive.

Fast Version

python main.py --path data/rnasolo764.txt --step 5000 --poststep 0 --batch_size 200 --worker_count 20 --repeat 5
python main.py --path data/eterna100.txt --step 5000 --poststep 0 --batch_size 50 --worker_count 20 --repeat 5
python main.py --path data/16s.txt --step 5000 --poststep 0 --batch_size 10 --worker_count 10 --repeat 5 --motif_path data/helix_motifs.txt --k_prune 20

This runs 5 repeated experiments and produces 5 CSV files plus a meta_data.json file under: results/output_timestamp/

Evaluation (without rival search):

python utils/metrics.py --meta results/output_timestamp/meta_data.json --eval

This will evaluate and output the metrics without rival search.

Boosting MFE-based metrics using rival search

python utils/metrics.py --meta results/output_timestamp/meta_data.json -r

This generates improved MFE and uMFE results and saves them to: results/output_timestamp/rival_serach_results.csv

Full Version

python main.py --path data/rnasolo764.txt --step 5000 --poststep 2500 --batch_size 200 --worker_count 20 --repeat 5
python main.py --path data/eterna100.txt --step 5000 --poststep 2500 --batch_size 50 --worker_count 20 --repeat 5

As with the fast version, this generates 5 CSV files and a meta_data.json file in: results/output_timestamp/

Evaluation (without rival search):

python utils/metrics.py --meta results/output_timestamp/meta_data.json --eval

This will evaluate and output the metrics without rival search.

Boosting MFE-based metrics using rival search

python utils/metrics.py --meta results/output_timestamp/meta_data.json -r

This generates improved MFE and uMFE results and saves them to: results/output_timestamp/rival_serach_results.csv